FirstRead / Technical appendix A4
The iteration loop
- Find a miss in the benchmark, and read why it was missed.
- Rewrite or add a question, or change a rule: they're code.
- Re-run the eval, and count what it cost in false alarms.
88.1% of real risk caught, with 1.5% of no-risk messages flagged.
- 1
Starting policy
Where we started: one cutoff per question, per level. 0c8e36b · 25 Sept
88.1% caught · 1.5% false alarms
- 2
Follow-up questions for borderline turns
Miss addressed: Weak signals that add up, with no one answer high enough on its own. c6e97c5 · 25 Sept
95.4% caught · 3.2% false alarms (+1.7)
- 3
A question about preparation
Miss addressed: Goodbye letters and gathering means reached CRISIS only by accident, or not at all. 5ddbd89 · 25 Sept
96.0% caught · 3.5% false alarms (+0.3)